CPU Tests: Microbenchmarks

Core-to-Core Latency

As the core count of modern CPUs is growing, we are reaching a time when the time to access each core from a different core is no longer a constant. Even before the advent of heterogeneous SoC designs, processors built on large rings or meshes can have different latencies to access the nearest core compared to the furthest core. This rings true especially in multi-socket server environments.

But modern CPUs, even desktop and consumer CPUs, can have variable access latency to get to another core. For example, in the first generation Threadripper CPUs, we had four chips on the package, each with 8 threads, and each with a different core-to-core latency depending on if it was on-die or off-die. This gets more complex with products like Lakefield, which has two different communication buses depending on which core is talking to which.

If you are a regular reader of AnandTech’s CPU reviews, you will recognize our Core-to-Core latency test. It’s a great way to show exactly how groups of cores are laid out on the silicon. This is a custom in-house test built by Andrei, and we know there are competing tests out there, but we feel ours is the most accurate to how quick an access between two cores can happen.

The core-to-core numbers are interesting, being worse (higher) than the previous generation across the board. Here we are seeing, mostly, 28-30 nanoseconds, compared to 18-24 nanoseconds with the 10700K. This is part of the L3 latency regression, as shown in our next tests.

One pair of threads here are very fast to access all cores, some 5 ns faster than any others, which again makes the layout more puzzling. 

Update 1: With microcode 0x34, we saw no update to the core-to-core latencies.

Cache-to-DRAM Latency

This is another in-house test built by Andrei, which showcases the access latency at all the points in the cache hierarchy for a single core. We start at 2 KiB, and probe the latency all the way through to 256 MB, which for most CPUs sits inside the DRAM (before you start saying 64-core TR has 256 MB of L3, it’s only 16 MB per core, so at 20 MB you are in DRAM).

Part of this test helps us understand the range of latencies for accessing a given level of cache, but also the transition between the cache levels gives insight into how different parts of the cache microarchitecture work, such as TLBs. As CPU microarchitects look at interesting and novel ways to design caches upon caches inside caches, this basic test proves to be very valuable.

Looking at the rough graph of the 11700K and the general boundaries of the cache hierarchies, we again see the changes of the microarchitecture that had first debuted in Intel’s Sunny Cove cores, such as the move from an L1D cache from 32KB to 48KB, as well as the doubling of the L2 cache from 256KB to 512KB.

The L3 cache on these parts look to be unchanged from a capacity perspective, featuring the same 16MB which is shared amongst the 8 cores of the chip.

On the DRAM side of things, we’re not seeing much change, albeit there is a small 2.1ns generational regression at the full random 128MB measurement point. We’re using identical RAM sticks at the same timings between the measurements here.

It’s to be noted that these slight regressions are also found across the cache hierarchies, with the new CPU, although it’s clocked slightly higher here, shows worse absolute latency than its predecessor, it’s also to be noted that AMD’s newest Zen3 based designs showcase also lower latency across the board.

With the new graph of the Core i7-11700K with microcode 0x34, the same cache structures are observed, however we are seeing better performance with L3.

The L1 cache structure is the same, and the L2 is of a similar latency. In our previous test, the L3 latency was 50.9 cycles, but with the new microcode is now at 45.1 cycles, and is now more in line with the L3 cache on Comet Lake.

Out at DRAM, our 128 MB point reduced from 82.4 nanoseconds to 72.8 nanoseconds, which is a 12% reduction, but not the +40% reduction that other media outlets are reporting as we feel our tools are more accurate. Similarly, for DRAM bandwidth, we are seeing a +12% memory bandwidth increase between 0x2C and 0x34, not the +50% bandwidth others are claiming. (BIOS 0x1B however, was significantly lower than this, resulting in a +50% bandwidth increase from 0x1B to 0x34.)

In the previous edition of our article, we questioned the previous L3 cycle being a larger than estimated regression. With the updated microcode, the smaller difference is still a regression, but more in line with our expectations. We are waiting to hear back from Intel what differences in the microcode encouraged this change.

Frequency Ramping

Both AMD and Intel over the past few years have introduced features to their processors that speed up the time from when a CPU moves from idle into a high powered state. The effect of this means that users can get peak performance quicker, but the biggest knock-on effect for this is with battery life in mobile devices, especially if a system can turbo up quick and turbo down quick, ensuring that it stays in the lowest and most efficient power state for as long as possible.

Intel’s technology is called SpeedShift, although SpeedShift was not enabled until Skylake.

One of the issues though with this technology is that sometimes the adjustments in frequency can be so fast, software cannot detect them. If the frequency is changing on the order of microseconds, but your software is only probing frequency in milliseconds (or seconds), then quick changes will be missed. Not only that, as an observer probing the frequency, you could be affecting the actual turbo performance. When the CPU is changing frequency, it essentially has to pause all compute while it aligns the frequency rate of the whole core.

We wrote an extensive review analysis piece on this, called ‘Reaching for Turbo: Aligning Perception with AMD’s Frequency Metrics’, due to an issue where users were not observing the peak turbo speeds for AMD’s processors.

We got around the issue by making the frequency probing the workload causing the turbo. The software is able to detect frequency adjustments on a microsecond scale, so we can see how well a system can get to those boost frequencies. Our Frequency Ramp tool has already been in use in a number of reviews.

Our ramp test shows a jump straight from 800 MHz up to 4900 MHz in around 17 milliseconds, or a frame at 60 Hz. 

Power Consumption: Hot Hot HOT CPU Tests: Office and Science
Comments Locked

541 Comments

View All Comments

  • Makste - Saturday, March 6, 2021 - link

    Yes. And we also got spoiled in the process. Now these are the kind of jumps we want with every release because AMD showed us that they were possible. The thrill of these performance jumps AMD has brought to the table are nothing short of exciting to the entire semiconductor realm. Things Intel should have been doing.
  • Qasar - Saturday, March 6, 2021 - link

    i read else where that the " reason " intel didnt go above quad cores, or increase performance more then 10% gen on gen before Zen, was cause if they did, then amd would of been put out of business.

    i would like to know is, when did/has intel ever done anything that was good for anyone other then intel ?
  • blppt - Sunday, March 7, 2021 - link

    Well, if AMD were to go under, then Intel would be subject to increased scrutiny of being a near-monopoly. Thats probably why they would have done such a thing (if this was true).

    Its similar to what Microsoft did way back in the day, propping up troubled Apple with cash infusions.
  • Qasar - Sunday, March 7, 2021 - link

    but the part is, NO one really knows what would of, or could of happened then. all it is, is one big IF.
  • Spunjji - Monday, March 8, 2021 - link

    That sounds like a rationalisation more than a reason - they had the clear motive of protecting the revenue from their "HEDT" and server processor lines. It was telling how quickly they dropped their trousers on those products after Ryzen launched.
  • Zizy - Monday, March 8, 2021 - link

    That would imply some incredibly hefty performance reserves to be unlocked when Intel felt threatened.
    Nah, all Intel seems to have done is to keep shrinking chip sizes at same 4C. This maybe was picked to help AMD, or just to help their bottom line. All clocks, IPC and whatever other improvements they had available were put in end products - otherwise these would be unlocked in Ice and Tiger lake, or this Rocket Lake abomination.
  • TheinsanegamerN - Monday, March 8, 2021 - link

    The biggest reason, and thisis one thing people dont want to admit, is there was simply no reason to. Their 6 core HDET parts were consistently slower then their 4 core peasant parts in consumer applications. So spending the money on making a new die just to have 6 cores on peasant class would have been a waste of money.
  • Cooe - Saturday, March 6, 2021 - link

    Zen 2 was NOT "Zen++". Not even close. It had a ton of MASSIVE changes vs Zen 1.

    Sure, Zen 2 is closer to Zen 1 than Zen 3 is to Zen 2, but that's not saying much at all....
  • Mighty Molecule - Friday, March 5, 2021 - link

    Remarkably underwhelming and disappointing. If the only distinction between 11900K and 10700K is left to binned core frequencies, you’re then left with the upgrades to PCIe 4 (which really is a designed delay from Comet Lake) and backported IPC’s at a tragic TDP. Yikes.
  • shabby - Friday, March 5, 2021 - link

    Ian can you guys start posting the actual wattage usage during each test beside the 125w rating? Would be nice seeing how much each cpu used during each specific test, if possible.

Log in

Don't have an account? Sign up now